AI Models
Which model each feature of CodeBook asks, what it actually ran, and how a platform administrator assigns one.
Admin → AI Models answers one question for the whole deployment: which model is used for what? Every feature of CodeBook that asks a model is listed by name, with a one-line description of what it does, how often it ran in the last 30 days and on which models, and a picker that chooses the model it asks from now on.
The list is not typed into the page by hand. It is generated from the feature registry in the code — the same registry every call site names itself with — so a feature that calls a model cannot exist without appearing here, and a feature that is removed disappears from the page with it.
What the page shows
The page is split by what the model does — chat models that read, decide or write, and embedding models that turn text into vectors for search — and within each, by the part of CodeBook the feature belongs to.
| Group | Features |
|---|---|
| Factory work | Agent run, Manager and worker chat, Crews, Planner, Workflow supervisor, Factory supervisor, Station evaluator, Context rerank |
| Judging and measurement | Scorer, Observer, Benchmark replay, Benchmark judge, Studio evaluation |
| Knowledge and the System Map | Domain extraction, System Map summaries, System Map documentation, Ask the System Map; and the embedding features Knowledge indexing, Knowledge search, System Map embeddings |
| Assistants | Concierge, Search answer, Client assistant |
Each card carries:
- the feature's name and what it does, in one sentence;
- its usage over the last 30 days — how many calls, how many tokens, and the share of calls each model answered (4 calls · 1.8k tokens · claude-sonnet-5-5 (75%), gpt-5.2 (25%)), or No calls in the last 30 days;
- the model picker, whose first option is always Default.
The usage line is what really ran, recorded by CodeBook at the moment of each call with the provider and model the call resolved to. It is not what the assignment says should run: when the two differ, that difference is the point of the line.
Default and assignments
Default means the model Bookbag stars for the platform — the one marked default in Platform AI at Bookbag, where the keys, the models themselves and their prices live. CodeBook never holds a key and the page never shows one.
Choosing a model on a card stores an assignment for that feature. From then on the feature asks that model, in this order of precedence:
- a model the caller named explicitly (a factory or station model policy, or a worker's pinned model) always wins;
- otherwise the feature's assignment, when the organization runs that model;
- otherwise the organization's default model.
The picker offers only models that fit the feature. A feature that uses tools — an agent at its station, the planner, the supervisors, the concierge — is offered only models that can call them; an embedding feature is offered only embedding models. Assigning a model the feature cannot use is refused by the server as well as hidden by the page.
An assignment is a platform-wide choice, so it applies only where the organization runs that model. Managed organizations run the platform's models and follow it directly. A self-hosted organization runs its own catalogue, and the assignment applies to it only when its catalogue contains a model with the same identifier; when it does not, the call runs on that organization's default, and the usage line records the model that really ran. Nothing here reaches into an organization's own keys.
Choosing Default again removes the assignment.
Where the models come from
The page reads the platform's model list from Bookbag. A Bookbag that exposes its platform scope answers directly with the starred model and each model's tool support. An older Bookbag is read through one of the deployment's managed organizations instead; when none exists, the page lists the models Bookbag offers and says that it could not learn which is starred or which can call tools, rather than guessing.
If Bookbag cannot be read at all, the page says so and assignments are refused until it can — an assignment is only ever checked against a model list that was actually read.
Who can use it
The page and both of its endpoints are platform-administrator only, with the same double check as every admin endpoint: authentication first, then the administrator check. A member of an organization, including its owner, is refused.
Usage records
Each call is recorded with its feature, provider, model, token counts and time, and nothing else — no prompt, no answer, no person. Records older than 90 days are pruned nightly. The 30-day figures on the page are built from these records, so they cover calls made since the page shipped.